Evaluating the Impact of Face Anonymization Methods on Computer Vision Tasks: A Trade-Off Between Privacy and Utility

Data anonymization is an essential prerequisite that enables data sharing in a privacy-preserving manner. However, anonymization affects the quality of the data and thus might affect the performance of later conducted data analysis. In this work, we investigate the influence of different image-based subject anonymization methods on the performance of the three common computer vision (CV) tasks: keypoint detection, instance segmentation, and face detection. We compare the anonymization methods on how much they affect the performance of the CV tasks, as well as their degree of anonymization. Furthermore, a re-identification attack scenario is applied. The study findings indicate a clear trade-off between increased privacy and decreased CV performance. Latent diffusion (called L-Diff in the following) shows the best results with minimal performance decrease on all CV tasks with a low recognition rate. However, the retrieval rate of our attack scenario, which measures the re-identification rate, remains high for all anonymization methods, which contrasts with the respective low level of recognition.